drivers/staging/erofs/utils.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * linux/drivers/staging/erofs/utils.c
   4  *
   5  * Copyright (C) 2018 HUAWEI, Inc.
   6  *             http://www.huawei.com/
   7  * Created by Gao Xiang <gaoxiang25@huawei.com>
   8  *
   9  * This file is subject to the terms and conditions of the GNU General Public
  10  * License.  See the file COPYING in the main directory of the Linux
  11  * distribution for more details.
  12  */
  13
  14 #include "internal.h"
  15 #include <linux/pagevec.h>
  16
  17 struct page *erofs_allocpage(struct list_head *pool, gfp_t gfp)
  18 {
  19         struct page *page;
  20
  21         if (!list_empty(pool)) {
  22                 page = lru_to_page(pool);
  23                 list_del(&page->lru);
  24         } else {
  25                 page = alloc_pages(gfp | __GFP_NOFAIL, 0);
  26
  27                 BUG_ON(page == NULL);
  28                 BUG_ON(page->mapping != NULL);
  29         }
  30         return page;
  31 }
  32
  33 /* global shrink count (for all mounted EROFS instances) */
  34 static atomic_long_t erofs_global_shrink_cnt;
  35
  36 #ifdef CONFIG_EROFS_FS_ZIP
  37
  38 /* radix_tree and the future XArray both don't use tagptr_t yet */
  39 struct erofs_workgroup *erofs_find_workgroup(
  40         struct super_block *sb, pgoff_t index, bool *tag)
  41 {
  42         struct erofs_sb_info *sbi = EROFS_SB(sb);
  43         struct erofs_workgroup *grp;
  44         int oldcount;
  45
  46 repeat:
  47         rcu_read_lock();
  48         grp = radix_tree_lookup(&sbi->workstn_tree, index);
  49         if (grp != NULL) {
  50                 *tag = radix_tree_exceptional_entry(grp);
  51                 grp = (void *)((unsigned long)grp &
  52                         ~RADIX_TREE_EXCEPTIONAL_ENTRY);
  53
  54                 if (erofs_workgroup_get(grp, &oldcount)) {
  55                         /* prefer to relax rcu read side */
  56                         rcu_read_unlock();
  57                         goto repeat;
  58                 }
  59
  60                 /* decrease refcount added by erofs_workgroup_put */
  61                 if (unlikely(oldcount == 1))
  62                         atomic_long_dec(&erofs_global_shrink_cnt);
  63                 BUG_ON(index != grp->index);
  64         }
  65         rcu_read_unlock();
  66         return grp;
  67 }
  68
  69 int erofs_register_workgroup(struct super_block *sb,
  70                              struct erofs_workgroup *grp,
  71                              bool tag)
  72 {
  73         struct erofs_sb_info *sbi;
  74         int err;
  75
  76         /* grp->refcount should not < 1 */
  77         BUG_ON(!atomic_read(&grp->refcount));
  78
  79         err = radix_tree_preload(GFP_NOFS);
  80         if (err)
  81                 return err;
  82
  83         sbi = EROFS_SB(sb);
  84         erofs_workstn_lock(sbi);
  85
  86         if (tag)
  87                 grp = (void *)((unsigned long)grp |
  88                         1UL << RADIX_TREE_EXCEPTIONAL_SHIFT);
  89
  90         err = radix_tree_insert(&sbi->workstn_tree,
  91                 grp->index, grp);
  92
  93         if (!err) {
  94                 __erofs_workgroup_get(grp);
  95         }
  96
  97         erofs_workstn_unlock(sbi);
  98         radix_tree_preload_end();
  99         return err;
 100 }
 101
 102 extern void erofs_workgroup_free_rcu(struct erofs_workgroup *grp);
 103
 104 int erofs_workgroup_put(struct erofs_workgroup *grp)
 105 {
 106         int count = atomic_dec_return(&grp->refcount);
 107
 108         if (count == 1)
 109                 atomic_long_inc(&erofs_global_shrink_cnt);
 110         else if (!count) {
 111                 atomic_long_dec(&erofs_global_shrink_cnt);
 112                 erofs_workgroup_free_rcu(grp);
 113         }
 114         return count;
 115 }
 116
 117 unsigned long erofs_shrink_workstation(struct erofs_sb_info *sbi,
 118                                        unsigned long nr_shrink,
 119                                        bool cleanup)
 120 {
 121         pgoff_t first_index = 0;
 122         void *batch[PAGEVEC_SIZE];
 123         unsigned freed = 0;
 124
 125         int i, found;
 126 repeat:
 127         erofs_workstn_lock(sbi);
 128
 129         found = radix_tree_gang_lookup(&sbi->workstn_tree,
 130                 batch, first_index, PAGEVEC_SIZE);
 131
 132         for (i = 0; i < found; ++i) {
 133                 int cnt;
 134                 struct erofs_workgroup *grp = (void *)
 135                         ((unsigned long)batch[i] &
 136                                 ~RADIX_TREE_EXCEPTIONAL_ENTRY);
 137
 138                 first_index = grp->index + 1;
 139
 140                 cnt = atomic_read(&grp->refcount);
 141                 BUG_ON(cnt <= 0);
 142
 143                 if (cleanup)
 144                         BUG_ON(cnt != 1);
 145
 146 #ifndef EROFS_FS_HAS_MANAGED_CACHE
 147                 else if (cnt > 1)
 148 #else
 149                 if (!erofs_workgroup_try_to_freeze(grp, 1))
 150 #endif
 151                         continue;
 152
 153                 if (radix_tree_delete(&sbi->workstn_tree,
 154                         grp->index) != grp) {
 155 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 156 skip:
 157                         erofs_workgroup_unfreeze(grp, 1);
 158 #endif
 159                         continue;
 160                 }
 161
 162 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 163                 if (erofs_try_to_free_all_cached_pages(sbi, grp))
 164                         goto skip;
 165
 166                 erofs_workgroup_unfreeze(grp, 1);
 167 #endif
 168                 /* (rarely) grabbed again when freeing */
 169                 erofs_workgroup_put(grp);
 170
 171                 ++freed;
 172                 if (unlikely(!--nr_shrink))
 173                         break;
 174         }
 175         erofs_workstn_unlock(sbi);
 176
 177         if (i && nr_shrink)
 178                 goto repeat;
 179         return freed;
 180 }
 181
 182 #endif
 183
 184 /* protected by 'erofs_sb_list_lock' */
 185 static unsigned int shrinker_run_no;
 186
 187 /* protects the mounted 'erofs_sb_list' */
 188 static DEFINE_SPINLOCK(erofs_sb_list_lock);
 189 static LIST_HEAD(erofs_sb_list);
 190
 191 void erofs_register_super(struct super_block *sb)
 192 {
 193         struct erofs_sb_info *sbi = EROFS_SB(sb);
 194
 195         mutex_init(&sbi->umount_mutex);
 196
 197         spin_lock(&erofs_sb_list_lock);
 198         list_add(&sbi->list, &erofs_sb_list);
 199         spin_unlock(&erofs_sb_list_lock);
 200 }
 201
 202 void erofs_unregister_super(struct super_block *sb)
 203 {
 204         spin_lock(&erofs_sb_list_lock);
 205         list_del(&EROFS_SB(sb)->list);
 206         spin_unlock(&erofs_sb_list_lock);
 207 }
 208
 209 unsigned long erofs_shrink_count(struct shrinker *shrink,
 210                                  struct shrink_control *sc)
 211 {
 212         return atomic_long_read(&erofs_global_shrink_cnt);
 213 }
 214
 215 unsigned long erofs_shrink_scan(struct shrinker *shrink,
 216                                 struct shrink_control *sc)
 217 {
 218         struct erofs_sb_info *sbi;
 219         struct list_head *p;
 220
 221         unsigned long nr = sc->nr_to_scan;
 222         unsigned int run_no;
 223         unsigned long freed = 0;
 224
 225         spin_lock(&erofs_sb_list_lock);
 226         do
 227                 run_no = ++shrinker_run_no;
 228         while (run_no == 0);
 229
 230         /* Iterate over all mounted superblocks and try to shrink them */
 231         p = erofs_sb_list.next;
 232         while (p != &erofs_sb_list) {
 233                 sbi = list_entry(p, struct erofs_sb_info, list);
 234
 235                 /*
 236                  * We move the ones we do to the end of the list, so we stop
 237                  * when we see one we have already done.
 238                  */
 239                 if (sbi->shrinker_run_no == run_no)
 240                         break;
 241
 242                 if (!mutex_trylock(&sbi->umount_mutex)) {
 243                         p = p->next;
 244                         continue;
 245                 }
 246
 247                 spin_unlock(&erofs_sb_list_lock);
 248                 sbi->shrinker_run_no = run_no;
 249
 250 #ifdef CONFIG_EROFS_FS_ZIP
 251                 freed += erofs_shrink_workstation(sbi, nr, false);
 252 #endif
 253
 254                 spin_lock(&erofs_sb_list_lock);
 255                 /* Get the next list element before we move this one */
 256                 p = p->next;
 257
 258                 /*
 259                  * Move this one to the end of the list to provide some
 260                  * fairness.
 261                  */
 262                 list_move_tail(&sbi->list, &erofs_sb_list);
 263                 mutex_unlock(&sbi->umount_mutex);
 264
 265                 if (freed >= nr)
 266                         break;
 267         }
 268         spin_unlock(&erofs_sb_list_lock);
 269         return freed;
 270 }
 271